mindarmour.natural_robustness.transform.image
This package include methods to generate natural perturbation samples.
- class mindarmour.natural_robustness.transform.image.Contrast(alpha=1, beta=0, auto_param=False)[source]
Contrast of an image.
- Parameters
Example
>>> img = cv2.imread('1.png') >>> img = np.array(img) >>> alpha = 0.1 >>> beta = 1 >>> trans = Contrast(alpha, beta) >>> dst = trans(img)
- class mindarmour.natural_robustness.transform.image.Curve(curves=3, depth=10, mode='vertical', auto_param=False)[source]
Curve picture using sin method.
- Parameters
curves (union[float, int]) – Divide width to curves of 2*math.pi, which means how many curve cycles. Suggested value range in [0.1. 5].
depth (union[float, int]) – Amplitude of sin method. Suggested value not exceed 1/10 of the length of the picture.
mode (str) – Direction of deformation. Optional value is ‘vertical’ or ‘horizontal’.
auto_param (bool) – Auto selected parameters. Selected parameters will preserve semantics of image.
Examples
>>> img = cv2.imread('x.png') >>> curves =1 >>> depth = 10 >>> trans = Curve(curves, depth, mode='vertical') >>> img_new = trans(img)
- class mindarmour.natural_robustness.transform.image.GaussianBlur(ksize=2, auto_param=False)[source]
Blurs the image using Gaussian blur filter.
- Parameters
Example
>>> img = cv2.imread('1.png') >>> img = np.array(img) >>> ksize = 5 >>> trans = GaussianBlur(ksize) >>> dst = trans(img)
- class mindarmour.natural_robustness.transform.image.GaussianNoise(factor=0.1, auto_param=False)[source]
Add gaussian noise of an image.
- Parameters
Example
>>> img = cv2.imread('1.png') >>> img = np.array(img) >>> factor = 0.1 >>> trans = GaussianNoise(factor) >>> dst = trans(img)
- class mindarmour.natural_robustness.transform.image.GradientBlur(point, kernel_num=3, center=True, auto_param=False)[source]
Gradient blur.
- Parameters
point (union[tuple, list]) – 2D coordinate of the Blur center point.
kernel_num (int) – Number of blur kernels. Suggested value range in [1, 8].
center (bool) – Blurred or clear at the center of a specified point.
auto_param (bool) – Auto selected parameters. Selected parameters will preserve semantics of image.
Example
>>> img = cv2.imread('xx.png') >>> img = np.array(img) >>> number = 5 >>> h, w = img.shape[:2] >>> point = (int(h / 5), int(w / 5)) >>> center = True >>> trans = GradientBlur(point, number, center) >>> new_img = trans(img)
- class mindarmour.natural_robustness.transform.image.GradientLuminance(color_start=(0, 0, 0), color_end=(255, 255, 255), start_point=(10, 10), scope=0.5, pattern='light', bright_rate=0.3, mode='circle', auto_param=False)[source]
Gradient adjusts the luminance of picture.
- Parameters
color_start (union[tuple, list]) – Color of gradient center. Default:(0, 0, 0).
color_end (union[tuple, list]) – Color of gradient edge. Default:(255, 255, 255).
start_point (union[tuple, list]) – 2D coordinate of gradient center.
scope (float) – Range of the gradient. A larger value indicates a larger gradient range. Default: 0.3.
pattern (str) – Dark or light, this value must be in [‘light’, ‘dark’].
bright_rate (float) – Control brightness. A larger value indicates a larger gradient range. If parameter ‘pattern’ is ‘light’, Suggested value range in [0.1, 0.7], if parameter ‘pattern’ is ‘dark’, Suggested value range in [0.1, 0.9].
mode (str) – Gradient mode, value must be in [‘circle’, ‘horizontal’, ‘vertical’].
auto_param (bool) – Auto selected parameters. Selected parameters will preserve semantics of image.
Examples
>>> img = cv2.imread('x.png') >>> height, width = img.shape[:2] >>> point = (height // 4, width // 2) >>> start = (255, 255, 255) >>> end = (0, 0, 0) >>> scope = 0.3 >>> pattern='light' >>> bright_rate = 0.3 >>> trans = GradientLuminance(start, end, point, scope, pattern, bright_rate, mode='circle') >>> img_new = trans(img)
- class mindarmour.natural_robustness.transform.image.MotionBlur(degree=5, angle=45, auto_param=False)[source]
Motion blur for a given image.
- Parameters
degree (int) – Degree of blur. This value must be positive. Suggested value range in [1, 15].
angle – (union[float, int]): Direction of motion blur. Angle=0 means up and down motion blur. Angle is counterclockwise.
auto_param (bool) – Auto selected parameters. Selected parameters will preserve semantics of image.
Example
>>> img = cv2.imread('1.png') >>> img = np.array(img) >>> angle = 0 >>> degree = 5 >>> trans = MotionBlur(degree=degree, angle=angle) >>> new_img = trans(img)
- class mindarmour.natural_robustness.transform.image.NaturalNoise(ratio=0.0002, k_x_range=(1, 5), k_y_range=(3, 25), auto_param=False)[source]
Add natural noise to an image.
- Parameters
ratio (float) – Noise density, the proportion of noise blocks per unit pixel area. Suggested value range in [0.00001, 0.001].
k_x_range (union[list, tuple]) – Value range of the noise block length.
k_y_range (union[list, tuple]) – Value range of the noise block width.
auto_param (bool) – Auto selected parameters. Selected parameters will preserve semantics of image.
Examples
>>> img = cv2.imread('xx.png') >>> img = np.array(img) >>> ratio = 0.0002 >>> k_x_range = (1, 5) >>> k_y_range = (3, 25) >>> trans = NaturalNoise(ratio, k_x_range, k_y_range) >>> new_img = trans(img)
- class mindarmour.natural_robustness.transform.image.Perspective(ori_pos, dst_pos, auto_param=False)[source]
Perform perspective transformation on a given picture.
- Parameters
Example
>>> img = cv2.imread('1.png') >>> img = np.array(img) >>> ori_pos = [[0, 0], [0, 800], [800, 0], [800, 800]] >>> dst_pos = [[50, 0], [0, 800], [780, 0], [800, 800]] >>> trans = Perspective(ori_pos, dst_pos) >>> dst = trans(img)
- class mindarmour.natural_robustness.transform.image.Rotate(angle=20, auto_param=False)[source]
Rotate an image of counter clockwise around its center.
- Parameters
Example
>>> img = cv2.imread('1.png') >>> img = np.array(img) >>> angle = 20 >>> trans = Rotate(angle) >>> dst = trans(img)
- class mindarmour.natural_robustness.transform.image.SaltAndPepperNoise(factor=0, auto_param=False)[source]
Add salt and pepper noise of an image.
- Parameters
Example
>>> img = cv2.imread('1.png') >>> img = np.array(img) >>> factor = 0.1 >>> trans = SaltAndPepperNoise(factor) >>> dst = trans(img)
- class mindarmour.natural_robustness.transform.image.Scale(factor_x=1, factor_y=1, auto_param=False)[source]
Scale an image in the middle.
- Parameters
factor_x (Union[float, int]) – Rescale in X-direction, x=factor_x*x. Suggested value range in [0.5, 1] and abs(factor_y - factor_x) < 0.5.
factor_y (Union[float, int]) – Rescale in Y-direction, y=factor_y*y. Suggested value range in [0.5, 1] and abs(factor_y - factor_x) < 0.5.
auto_param (bool) – Auto selected parameters. Selected parameters will preserve semantics of image.
Example
>>> img = cv2.imread('1.png') >>> img = np.array(img) >>> factor_x = 0.7 >>> factor_y = 0.6 >>> trans = Scale(factor_x, factor_y) >>> dst = trans(img)
- class mindarmour.natural_robustness.transform.image.Shear(factor=0.2, direction='horizontal', auto_param=False)[source]
Shear an image, for each pixel (x, y) in the sheared image, the new value is taken from a position (x+factor_x*y, factor_y*x+y) in the origin image. Then the sheared image will be rescaled to fit original size.
- Parameters
Example
>>> img = cv2.imread('1.png') >>> img = np.array(img) >>> factor = 0.2 >>> trans = Shear(factor, direction='horizontal') >>> dst = trans(img)
- class mindarmour.natural_robustness.transform.image.Translate(x_bias=0, y_bias=0, auto_param=False)[source]
Translate an image.
- Parameters
x_bias (Union[int, float]) – X-direction translation, x = x + x_bias*image_width. Suggested value range in [-0.1, 0.1].
y_bias (Union[int, float]) – Y-direction translation, y = y + y_bias*image_length. Suggested value range in [-0.1, 0.1].
auto_param (bool) – Auto selected parameters. Selected parameters will preserve semantics of image.
Example
>>> img = cv2.imread('1.png') >>> img = np.array(img) >>> x_bias = 0.1 >>> y_bias = 0.1 >>> trans = Translate(x_bias, y_bias) >>> dst = trans(img)